JP7676578B2 - Method, apparatus and computer program for content-adaptive online training for end-to-end (E2E) neural image compression (NIC) using neural networks - Google Patents

Description

CROSS-REFERENCE TO RELATED APPLICATIONS This application is based on and claims priority to U.S. Provisional Patent Application No. 63/289,044, filed December 13, 2021, and U.S. Patent Application No. 17/969,242, filed October 19, 2022, the disclosures of which are incorporated by reference in their entireties herein.

Hybrid video codecs can be difficult to optimize as a whole. Improvements in a single module may not result in coding gains in overall performance. In recent years, standards groups and companies have been actively investigating the potential demand for standardization of future video coding technologies. These standards groups and companies have established the JPEG-AI group, which focuses on AI-based end-to-end neural image compression using deep neural networks (DNNs). And the Chinese Audio-Video Coding (AVS) standard has also formed an AVS-AI special group to work on neural image and video compression technologies. The success of recent methods has led to more and more industrial interest in advanced neural image and video compression methodologies.

However, in the prior art, neural network-based video or image coding frameworks are limited to a specific type of compression framework. To accommodate various types of frameworks, the prior art systems may require increased computing memory/cost and increased rate-distortion loss, leading to a decrease in performance of the entire image or video framework/process.

Therefore, there is a need for a way to optimize the coding framework and improve overall performance.

According to an embodiment, a method for content-adaptive online training for multiple blocks in neural image compression is provided.

According to an aspect of the present disclosure, a method for content-adaptive online training for end-to-end (E2E) neural image compression (NIC) based on several patterns using a neural network, executed by at least one processor, is provided. The method includes receiving an input image to an E2E NIC framework, dividing the input image into a number of blocks, selecting a subset of blocks from the number of blocks, the subset of blocks sharing the same pattern, preprocessing a neural network of the E2E NIC framework, the preprocessed neural network being applied to the selected subset of blocks, calculating update parameters using the preprocessed neural network, and generating an updated E2E NIC framework based on the update parameters.

The method may further include encoding the plurality of blocks and the update parameters to generate a compressed representation of the plurality of blocks and a compressed representation of the update parameters, decoding the compressed representation of the update parameters to generate decoded update parameters, updating the E2E NIC framework based on the decoded update parameters, and decoding the compressed representation of the plurality of blocks based on the updated E2E NIC framework to generate a reconstructed image.

The method may further include determining a distortion loss of the reconstructed image based on the consumption of the compressed representations of the plurality of blocks and the update parameters, the trade-off hyperparameter, and the distortion between the block residuals of the compressed representations of the plurality of blocks and the block residuals of the decoded compressed representations of the plurality of blocks.

In some embodiments, the same pattern is determined based on the RGB variance of multiple blocks, or the YUV variance of multiple blocks.

In some embodiments, the update parameters include a learning rate and a number of steps, where the learning rate and the number of steps are selected based on characteristics of the input image.

In some embodiments, the characteristic of the input image is one of the RGB variance of the input image and the RD performance of the input image.

The method may further include, when preprocessing the neural network, fine-tuning the neural network using a plurality of blocks.

According to another aspect of the present disclosure, there is provided an apparatus for a pattern-based content E2E NIC using a neural network, comprising at least one memory configured to store computer program code and at least one processor configured to read the computer program code and act as instructed by the computer program code. The computer program code includes a receiving code configured to cause at least one processor to receive an input image to the E2E NIC framework; a dividing code configured to cause the at least one processor to divide the input image into a plurality of blocks; a selection code configured to cause the at least one processor to select a subset of blocks from the plurality of blocks, the subset of blocks sharing a same pattern; a preprocessing code configured to cause the at least one processor to preprocess a neural network of the E2E NIC framework, the preprocessed neural network being applied to the selected subset of blocks; a computing code configured to cause the at least one processor to calculate update parameters using the preprocessed neural network; and a generating code configured to cause the at least one processor to generate an updated E2E NIC framework based on the update parameters.

The apparatus may further include an encoding code configured to cause the at least one processor to encode the plurality of blocks and the update parameters to generate a compressed representation of the plurality of blocks and a compressed representation of the update parameters; a first decoding code configured to cause the at least one processor to decode the compressed representation of the update parameters to generate decoded update parameters; an update code configured to cause the at least one processor to update the E2E NIC framework based on the decoded update parameters; and a second decoding code configured to cause the at least one processor to decode the compressed representation of the plurality of blocks based on the updated E2E NIC framework to generate a reconstructed image.

The apparatus may further include distortion loss determination code that causes at least one processor to determine a distortion loss of the reconstructed image based on the consumption of the compressed representations of the plurality of blocks and update parameters, the trade-off hyperparameters, and the distortion between the block residuals of the compressed representations of the plurality of blocks and the block residuals of the decoded compressed representations of the plurality of blocks.

According to another aspect of the present disclosure, a non-transitory computer-readable medium is provided that stores instructions executed by at least one processor of an apparatus for content-adaptive online training for E2E NICs based on a number of patterns using a neural network. The instructions cause the at least one processor to receive an input image to an E2E NIC framework, divide the input image into a number of blocks, select a subset of the blocks from the number of blocks, the subset of blocks sharing the same pattern, preprocess a neural network of the E2E NIC framework, apply the preprocessed neural network to the selected subset of blocks, calculate update parameters using the preprocessed neural network, and generate an updated E2E NIC framework based on the update parameters.

The non-transitory computer-readable medium may further include instructions to cause at least one processor to encode the plurality of blocks and the update parameters to generate a compressed representation of the plurality of blocks and a compressed representation of the update parameters, decode the compressed representation of the update parameters to generate decoded update parameters, update the E2E NIC framework based on the decoded update parameters, and decode the compressed representation of the plurality of blocks based on the updated E2E NIC framework to generate a reconstructed image.

The non-transitory computer-readable medium may further include instructions for causing at least one processor to determine a distortion loss of the reconstructed image based on consumption of the compressed representations of the plurality of blocks and update parameters, trade-off hyperparameters, and distortion between block residuals of the compressed representations of the plurality of blocks and block residuals of the decoded compressed representations of the plurality of blocks.

Additional embodiments are set forth in the description that follows and, in part, will be apparent from the description and/or may be learned by practice of the presented embodiments of the present disclosure.

1 is a flowchart of an overview of a content-adaptive online training process for end-to-end (E2E) neural image compression (NIC), according to an embodiment. FIG. 1 illustrates an environment in which the methods, apparatus, and systems described herein may be implemented, according to an embodiment. 3 is a block diagram illustrating example components of one or more devices of FIG. 2. FIG. 2 is a diagram showing an example of block-based image coding. 1 is an example showing content-adaptive online training of multiple blocks based on several patterns. 1 is a flowchart of an example coding process, according to an embodiment. FIG. 1 illustrates an exemplary block diagram illustrating an end-to-end (E2E) neural image compression (NIC) framework with content-adaptive online training, according to an embodiment. 1 is a flow chart illustrating a method for content-adaptive online training for end-to-end (E2E) neural image compression (NIC) based on several patterns using neural networks, according to an embodiment. FIG. 1 is a block diagram of an example of computer code for content-adaptive online training for end-to-end (E2E) neural image compression (NIC) based on several patterns using neural networks, according to an embodiment.

The following detailed description of the exemplary embodiments refers to the accompanying drawings. The same reference numbers in different drawings may identify the same or similar elements.

The foregoing disclosure provides illustrations and descriptions, but is not intended to be exhaustive or to limit the implementation to the precise form disclosed. Modifications and variations are possible in light of the above disclosure or may be acquired from practice of the implementations. Moreover, one or more features or components of one embodiment may be incorporated or combined with another embodiment (or one or more features of another embodiment). In addition, it will be understood that in the flowcharts and descriptions of operations presented below, one or more operations may be omitted, one or more operations may be added, one or more operations may be performed (at least partially) simultaneously, and the order of one or more operations may be changed.

It will be apparent that the systems and/or methods described herein may be implemented in different forms of hardware, software, or a combination of hardware and software. The actual dedicated control hardware or software code used to implement these systems and/or methods is not intended to limit the implementation. Thus, the operation and behavior of the systems and/or methods have been described herein without reference to any specific software code. It should be understood that software and hardware may be designed to implement the systems and/or methods based on the description herein.

Although certain combinations of features are recited in the claims and/or disclosed herein, these combinations are not intended to limit the disclosure of possible implementations. Indeed, many of these features may be combined in ways not specifically recited in the claims and/or disclosed herein. Although each dependent claim listed below may depend directly on only one claim, the disclosure of possible implementations includes each dependent claim in combination with every other claim in the claim set.

The proposed features described below may be used separately or combined in any order. Furthermore, the embodiments may be implemented by processing circuitry (e.g., one or more processors or one or more integrated circuits). In one example, one or more processors execute a program stored on a non-transitory computer-readable medium.

No element, act, or instruction used herein should be construed as critical or essential unless expressly described as such. Also, as used herein, the articles "a" and "an" are intended to include one or more items and may be used interchangeably with "one or more." When only one item is intended, the term "one" or similar words are used. Also, as used herein, the terms "has," "have," "having," "include," "including," and the like are intended to be open-ended terms. Furthermore, the phrase "based on" is intended to mean "based at least in part on," unless otherwise specified. Furthermore, phrases such as "at least one of [A] and [B]," or "at least one of [A] or [B]," should be understood to include only A, only B, or both A and B.

Exemplary embodiments of the present disclosure provide a method and apparatus for content-adaptive online training of multiple blocks in an optimized network of end-to-end (E2E) neural image compression (NIC). The E2E optimized network may be, for example, an artificial neural network (ANN)-based image coding framework. In the ANN-based video coding framework, different modules can be jointly optimized from input to output to improve the final objective (e.g., rate-distortion performance) by performing a machine learning process, resulting in an E2E optimized NIC.

Figure 1 is a flowchart of an overview of a content-adaptive online training process for an E2E NIC, as performed by a content-adaptive online training NIC framework, a content-adaptive online training system, etc., according to an embodiment.

First, an input image (or video sequence) is received (S110). Next, in S120, the image is divided into a number of blocks. Block-wise image coding may be performed on the divided blocks to compress the blocks. In S130, a set of blocks is selected from the number of blocks. The selection is made based on a pattern determined to be present in the number of blocks. In S140, pre-processing of the content-adaptive online training NIC framework is performed and applied to the selected set of blocks to fine-tune the network. In S150, update parameters are generated based on the pre-processed (i.e., fine-tuned) network. The update parameters may include, for example, but are not limited to, a step size (i.e., learning rate) and a number of steps. The blocks and the generated update parameters are then encoded, such as by a DNN encoder, and then decoded, such as by a DNN decoder (S160). The decoded update parameters are used to update the NIC framework (S170). Finally, the decoder of the updated NIC framework is used to decode and generate a final image. That is, in S180, a reconstructed image is generated based on the updated NIC framework.

FIG. 2 is a diagram of an environment 200 in which the methods, apparatus, and systems described herein may be implemented, according to an embodiment.

As shown in FIG. 2, environment 200 may include user devices 210, platform 220, and network 230. The devices of environment 200 may be interconnected via wired connections, wireless connections, or a combination of wired and wireless connections.

User device 210 includes one or more devices capable of receiving, generating, storing, processing, and/or providing information associated with platform 220. For example, user device 210 may include a computing device (e.g., a desktop computer, a laptop computer, a tablet computer, a handheld computer, a smart speaker, a server, etc.), a mobile phone (e.g., a smartphone, a wireless phone, etc.), a wearable device (e.g., smart glasses or a smart watch), or a similar device. In some implementations, user device 210 can receive information from and/or transmit information to platform 220.

Platform 220 includes one or more devices as described elsewhere herein. In some implementations, platform 220 may include a cloud server or a collection of cloud servers. In some implementations, platform 220 may be designed to be modular such that software components may be swapped in or out. Thus, platform 220 may be easily and/or quickly reconfigured for different uses.

In some implementations, as shown, platform 220 may be hosted within cloud computing environment 222. Notably, although the implementations described herein describe platform 220 as being hosted within cloud computing environment 222, in some implementations platform 220 may not be cloud-based (i.e., may be implemented outside of a cloud computing environment) or may be partially cloud-based.

Cloud computing environment 222 includes an environment that hosts platform 220. Cloud computing environment 222 may provide computational services, software services, data access services, storage services, and the like that do not require awareness by an end user (e.g., user device 210) of the physical location and configuration of one or more systems and/or one or more devices that host platform 220. As illustrated, cloud computing environment 222 may include a group of computing resources 224 (collectively referred to as "computing resources 224" and individually referred to as "computing resource 224").

Computing resources 224 include one or more personal computers, workstation computers, server devices, or other types of computing and/or communication devices. In some implementations, computing resources 224 may host platform 220. Cloud resources may include compute instances running within computing resources 224, storage devices provided within computing resources 224, data transfer devices provided by computing resources 224, and the like. In some implementations, computing resources 224 may communicate with other computing resources 224 via wired connections, wireless connections, or a combination of wired and wireless connections.

As further shown in FIG. 2, computing resources 224 may include a group of cloud resources, such as one or more applications ("APPs") 224-1, one or more virtual machines ("VMs") 224-2, one or more virtualized storage ("VS") 224-3, or one or more hypervisors ("HYPs") 224-4.

Application 224-1 includes one or more software applications that can be provided to or accessed by user device 210 and/or platform 220. Application 224-1 may eliminate the need to install and run software applications on user device 210. For example, application 224-1 may include software associated with platform 220 and/or any other software that can be provided via cloud computing environment 222. In some implementations, one application 224-1 may send and receive information to one or more other applications 224-1 via virtual machine 224-2.

Virtual machine 224-2 includes a software implementation of a machine (e.g., a computer) that executes programs like a physical machine. Virtual machine 224-2 may be either a system virtual machine or a process virtual machine, depending on the use by virtual machine 224-2 and its correspondence to any real machine. A system virtual machine can provide a complete system platform that supports the execution of a complete operating system ("OS"). A process virtual machine may execute a single program and support a single process. In some implementations, virtual machine 224-2 may run on behalf of a user (e.g., user device 210) and manage the infrastructure of cloud computing environment 222, such as data management, synchronization, or long-term data transfer.

Virtualized storage 224-3 includes one or more storage systems and/or one or more devices that use virtualization techniques within the storage systems or devices of computing resources 224. In some implementations, within the context of a storage system, types of virtualization may include block virtualization and file virtualization. Block virtualization may refer to the abstraction (or separation) of logical storage from physical storage such that the storage system may be accessed without regard to the physical storage or heterogeneous structure. The separation may allow flexibility in how storage system administrators manage storage for end users. File virtualization may eliminate dependencies between data accessed at the file level and where the file is physically stored. This may enable optimization of storage usage, server consolidation, and/or performing non-disruptive file migrations.

Hypervisor 224-4 may provide hardware virtualization technology that allows multiple operating systems (e.g., "guest operating systems") to run simultaneously on a host computer, such as computing resource 224. Hypervisor 224-4 may present a virtual operating platform to the guest operating systems and may manage the execution of the guest operating systems. Multiple instances of various operating systems may share virtualized hardware resources.

Network 230 may include one or more wired and/or wireless networks. For example, network 230 may include a cellular network (e.g., a fifth generation (5G) network, a long term evolution (LTE) network, a third generation (3G) network, a code division multiple access (CDMA) network, etc.), a public land mobile network (PLMN), a local area network (LAN), a wide area network (WAN), a metropolitan area network (MAN), a telephone network (e.g., a public switched telephone network (PSTN)), a private network, an ad-hoc network, an intranet, the Internet, an optical fiber-based network, etc., and/or a combination of these or other types of networks.

The number and arrangement of devices and networks shown in FIG. 2 are provided as an example. In practice, there may be additional devices and/or networks, fewer devices and/or networks, different devices and/or networks, or different arrangements of devices and/or networks than those shown in FIG. 2. Furthermore, two or more devices shown in FIG. 2 may be implemented within a single device, or a single device shown in FIG. 2 may be implemented as multiple distributed devices. Additionally or alternatively, a set of devices (e.g., one or more devices) of environment 200 may perform one or more functions described as being performed by another set of devices of environment 200.

Figure 3 is a block diagram illustrating example components of one or more devices of Figure 2.

The device 300 may correspond to the user device 210 and/or the platform 220. As shown in FIG. 3, the device 300 may include a bus 310, a processor 320, a memory 330, a storage component 340, an input component 350, an output component 360, and a communication interface 370.

The bus 310 includes components that enable communication between the components of the device 300. The processor 320 is implemented in hardware, software, or a combination of hardware and software. The processor 320 is a central processing unit (CPU), a graphics processing unit (GPU), an accelerated processing unit (APU), a microprocessor, a microcontroller, a digital signal processor (DSP), a field programmable gate array (FPGA), an application specific integrated circuit (ASIC), or another type of processing component. In some implementations, the processor 320 includes one or more processors that can be programmed to perform functions. The memory 330 includes random access memory (RAM), read only memory (ROM), and/or another type of dynamic or static storage device (e.g., flash memory, magnetic memory, and/or optical memory) that stores information and/or instructions used by the processor 320.

Storage component 340 stores information and/or software related to the operation and use of device 300. For example, storage component 340 may include a hard disk (e.g., a magnetic disk, an optical disk, a magneto-optical disk, and/or a solid-state disk), a compact disk (CD), a digital versatile disk (DVD), a floppy disk, a cartridge, a magnetic tape, and/or another type of non-transitory computer-readable medium along with a corresponding drive.

The input components 350 include components (e.g., a touch screen display, a keyboard, a keypad, a mouse, buttons, switches, and/or a microphone) that enable the device 300 to receive information, such as through user input. Additionally or alternatively, the input components 350 may include sensors for sensing information (e.g., a global positioning system (GPS) component, an accelerometer, a gyroscope, and/or an actuator). The output components 360 include components (e.g., a display, a speaker, and/or one or more light emitting diodes (LEDs)) that provide output information from the device 300.

The communication interface 370 includes transceiver-like components (e.g., a transceiver and/or a separate receiver and transmitter) that enable the device 300 to communicate with other devices, such as via a wired connection, a wireless connection, or a combination of wired and wireless connections. The communication interface 370 may enable the device 300 to receive information from another device and/or provide information to another device. For example, the communication interface 370 may include an Ethernet interface, an optical interface, a coaxial interface, an infrared interface, a radio frequency (RF) interface, a universal serial bus (USB) interface, a Wi-Fi interface, a cellular network interface, etc.

Device 300 may perform one or more of the processes described herein. Device 300 may perform these processes in response to processor 320 executing software instructions stored by a non-transitory computer-readable medium, such as memory 330 and/or storage component 340. A computer-readable medium is defined herein as a non-transitory memory device. A memory device includes memory space within a single physical storage device or memory space scattered across multiple physical storage devices.

Software instructions may be loaded into memory 330 and/or storage component 340 from another computer-readable medium or from another device via communication interface 370. The software instructions stored in memory 330 and/or storage component 340, when executed, may cause processor 320 to perform one or more processes described herein. Additionally or alternatively, hardwired circuitry may be used in place of or in combination with software instructions to perform one or more processes described herein. Thus, the implementations described herein are not limited to any particular combination of hardware circuitry and software.

The number and arrangement of components shown in FIG. 3 are provided as an example. In practice, device 300 may include additional, fewer, different, or differently arranged components than those shown in FIG. 3. Additionally or alternatively, a set of components (e.g., one or more components) of device 300 may perform one or more functions described as being performed by another set of components of device 300.

In an embodiment, any one of the operations or processes of Figures 4 through 9 may be implemented by or using any one of the elements shown in Figures 2 and 3.

According to some embodiments, the general process of neural network based image compression may be as follows: Given an image or video sequence x, the goal of the NIC is to generate a compressed representation x that is compact for storage and transmission purposes.
The solution is to use the image x as input to a DNN encoder to compute the compressed representation
is used as the input to the DNN decoder to generate the image
Some NIC methods can take a variational autoencoder (VAE) structure, where a DNN encoder directly uses the entire image x as its input, which is passed through a set of network layers that act like a black box, to produce an output representation (i.e., a compressed representation
Correspondingly, the DNN decoder computes the compressed representation
It takes the whole as its input, which goes through another set of network layers that act like another black box, reconstructing the image
The rate-distortion (RD) loss is calculated by using the following objective loss function L with a trade-off hyperparameter λ to compute the reconstructed image
Distortion loss
and compressed representation
The bit consumption R of the BER is optimized to achieve a trade-off between the bit consumption R of the BER.

The embodiment relates to a content-adaptive E2E online training NIC framework for multiple blocks. In content-adaptive online training, to train multiple blocks, an input image x is first split into multiple blocks. For example, an image of size 2048x2048 may be split into 16 blocks of size 512x512. A more detailed description is provided with reference to Figs. 4 and 5. Splitting the input image saves computing memory and enables parallel computing. Each block is then compressed individually by a rate-distortion compression method (such as the general NIC process described). In particular, the split input image is converted into a compressed representation of the input image x,
The compressed representation
Then the output (i.e. the reconstructed image
) is input to a decoder (e.g., a DNN decoder) to generate a post-augmented network. A neural network-based pre-trained (or not pre-trained) network is used as the post-augmented network. The online training of this post-augmented network is based on multiple blocks. During the online training, partial (or overall) parameters of the post-augmented network may be updated. The updated partial (or overall) parameters are encoded into a bitstream together with the encoded multiple blocks. Using these updated parameters (fine-tuned by one or more images), the DNN decoder may decode the encoded blocks and achieve good compression performance. This fine-tuning process is used as a pre-processing step in the content-adaptive online training NIC framework to enhance the compression performance of any pre-trained E2E NIC compression method. In some embodiments, this method may also be used in the E2E NIC framework, trained by the content-adaptive online training NIC framework itself.

A detailed description of pre-processing of the content-adaptive online training NIC framework in accordance with one or more embodiments is provided herein.

As described above, the pre-trained post-augmentation network receives a number of blocks as input. The post-augmentation network is then fine-tuned based on the input number of blocks. The fine-tuned post-augmentation network is used to obtain update parameters, which update the NIC framework. In this manner, the post-augmentation network can be applied to the target image content. When fine-tuning the post-augmentation network, one or more of the network parameters can be updated.

In some embodiments, the parameters may be updated in whole or in part. For example, the parameters may be updated in only a portion of the post-augmentation network (such as the last layer of a convolutional neural network). As another example, the parameters may be updated in multiple or all modules of the post-augmentation network.

In some embodiments, only the bias terms are optimized and updated. In another exemplary embodiment, the coefficient (weight) terms are optimized. Alternatively, for example, all parameters may be optimized.

In some embodiments, the post-augmentation network is fine-tuned to generate an updated NIC framework based on a number of blocks that form a large block/image. In some embodiments, the post-augmentation network is fine-tuned and the fine-tuned post-augmentation network generates an updated framework based on a set of blocks that may not be adjacent to each other. For example, the set of blocks may be randomly selected from the input number of blocks.

The fine-tuning process involves multiple epochs during which parameters are updated in this iterative online training process. When the training loss (e.g., as determined based on the target loss function in Equation 1) flattens or is about to flatten, the fine-tuning is stopped. There are two key hyperparameters in the content-adaptive online training NIC framework: step size and number of steps. The step size indicates the "learning rate" of the online training NIC framework. Images with different types of content may correspond to different step sizes to obtain the best optimization results. The number of steps indicates the number of updates manipulated. Along with the target loss function (Equation 1), the hyperparameters are used in the online learning process. For example, the step size may be used in the gradient descent algorithm or backpropagation calculations performed in the learning process. The number of iterations may be used as a threshold for the maximum number of iterations to control when the learning process may be terminated. In some embodiments, the learning rate (i.e., step size) may be changed at each step by a scheduler during the iterative online training process. The scheduler determines the learning rate value, which may increase, decrease, or remain the same for several intervals. There may be a single scheduler or multiple (different) schedulers for different input images. Multiple update parameters may be generated based on multiple learning rate schedulers, and a scheduler with good compression performance may be selected for each of the update parameters. At the end of the fine-tuning process, the update parameters are calculated. In some embodiments, the update parameters are then compressed at the end of the fine-tuning process. For example, a compression algorithm (such as LZMA2) may be used to compress the update parameters. In another exemplary embodiment, no compression of the update parameters is performed.

In some embodiments, the update parameters are calculated as the difference between the fine-tuned parameters and the pre-trained parameters. In some embodiments, the update parameters are the fine-tuned parameters. In another exemplary embodiment, the update parameters are some transform of the fine-tuned parameters.

Figure 4 illustrates an example of block-based image coding. In an embodiment of the content-adaptive online training NIC framework, a block-based coding scheme is used to compress image frames instead of directly encoding the entire input image. With a block-based coding scheme, the entire input image is first divided into blocks of equal (or different) sizes, and the blocks are compressed separately.

As shown in FIG. 4, the image 400 may first be divided into blocks (indicated by dashed lines in FIG. 4) and the divided blocks may be compressed instead of the image 400 itself. In FIG. 4, the compressed blocks are shaded and the blocks to be compressed are not shaded. The divided blocks may be of equal or unequal size. The step size per block may be different. To this end, different step sizes may be assigned to the image 400 to achieve better compression results. Block 410 is an example of one of the divided blocks having height h and width w. The blocks go through a block-wise image coding process to generate a bitstream of coded information.

FIG. 5 is an example illustrating content-adaptive online training of multiple blocks according to an embodiment. As shown in FIG. 5, an input image x may be divided into blocks. The size of the blocks and the content-adaptive regions may be different and are not limited by the example of FIG. 5. For example, the blocks may be the same or different sizes. Then, the pattern of each block is determined. A subset of blocks (including one or more blocks) having the same pattern may be selected. One or more sets of blocks may be selected and used within a single image to train the NIC framework. In some embodiments, one or more sets of blocks are selected from one or more input images or sequence videos. The selected subset of blocks is then used to generate a reconstructed image
The training data are processed in the content-adaptive online training NIC framework to generate

As shown in FIG. 5, a subset of blocks 510 is selected based on having the same pattern and used to train the NIC framework. The NIC framework trained based on the subset of blocks 510 is then used to train the reconstructed image
Another subset of blocks 520 is selected based on having the same pattern, and then used to generate the reconstructed image
The reconstructed image is then processed to produce
and reconstructed image
may be the output image before or after the filtering process.

The patterns (or pattern types) are not limited by the embodiment. The various attributes that may classify a block may be the patterns used for blocks of the same pattern selected for the content-adaptive training process.

In some embodiments, the pattern is based on the red-green-blue (RGB) distribution of the blocks. In another exemplary embodiment, the pattern is based on the luma, red projection, and blue projection model (i.e., YUV model) distribution of the blocks. In yet another exemplary embodiment, the pattern is based on the dominant colors present in one or more blocks.

The step size (i.e., the learning rate of the content-adaptive online training NIC framework) may be selected based on image (or block) characteristics. For example, the image characteristics may be based on the RGB color model and RGB variance of the image. Furthermore, in some embodiments, the step size may be selected based on the RD performance of the image (or block). Thus, according to the embodiment, multiple update parameters may be generated based on different step sizes, and for each update parameter, a step size with good compression performance may be selected.

To achieve better compression results, multiple learning rate schedules may be assigned to different blocks. In some embodiments, all blocks share the same learning rate schedule. The selection of the learning rate scheduler may be based on characteristics of the block, such as the RGB variance of the block or the RD performance of the block.

Different blocks may update different parameters in different modules (e.g., in the context module or the hyperdecoder) or different types of parameters (biases or weights) of the content-adaptive online training NIC framework according to the embodiment. In some embodiments, all blocks share the same update parameters. The (updated) parameters may be selected based on characteristics of the block, such as the RGB variance of the block or the RD performance of the block.

Different blocks may choose different ways to transform the update parameters. For example, in some embodiments, one block may choose to update parameters based on the difference between fine-tuned parameters and pre-trained parameters. Another block may choose to update parameters directly. In some embodiments, all blocks updated parameters in the same way. The method of transforming the update parameters may be selected based on characteristics of the block, such as the RGB variance of the block or the RD performance of the block.

Different blocks may select different methods to compress the update parameters. For example, one block may use the LZMA2 algorithm to compress the update parameters. Another block may use the bzip2 algorithm to compress the update parameters. The embodiments are not limited in this respect and may use any suitable compression algorithm to compress the parameters. In some embodiments, all blocks use the same method to compress (or not compress) the update parameters. The compression method may be selected based on characteristics of the block, such as the RGB variance of the block or the RD performance of the block.

The coding process of the content-adaptive online training NIC framework that is applied to the blocks to generate a reconstructed image is described with reference to FIG. 6, which is a flow chart of an example of the coding process, according to an embodiment.

First, in S610, the NIC framework encodes the input image and update parameters. Then, the encoded input and update parameters are decoded (S620). If the update parameters have been compressed (S630 is "yes"), the update parameters obtained from the online training process are first decompressed (S640). If the update parameters have not been compressed (S630 is "no"), the process proceeds to S650. In S650, the NIC framework is updated on the decoder side using the decoded update parameters from S620 or the decompressed decoded update parameters from S640. Finally, in S660, the (reconstructed image
The updated NIC framework decoder (to generate {overscore (x)}) is used for image decoding. Based on how the parameters were transformed, the updated parameter values update the original (pre-trained) bias terms.

The embodiments do not impose any limitations on the methods used for, for example, the neural encoder, the encoder, the decoder, and the neural decoder. The content-adaptive online training method according to the embodiments may be adapted to different types of NIC frameworks. For example, the process may be performed using different types of encoding and decoding DNNs.

Figure 7 is an example block diagram 700 of an E2E NIC framework using content-adaptive online training, according to an embodiment.

As shown in FIG. 7, the E2E NIC framework includes a primary encoder 710, a primary decoder 720, a hyperencoder 730, a hyperdecoder 740, and a context model 750. The E2E NIC framework may include one or more such modules. The E2E NIC framework further includes a quantizer 760/761, an arithmetic coder 770/771, and an arithmetic decoder 780/781. The same or similar modules are represented by the same reference numbers. The E2E NIC framework may include one or more modules not shown in FIG. 7.

The E2E NIC framework may use any DNN-based image compression method, such as the scaled hyperprior encoder-decoder framework (or Gaussian mixture likelihood framework) and its variants, RNN-based recursive compression methods and their variants.

According to an embodiment of the present disclosure, the E2E NIC framework may utilize block diagram 700 as follows: Given an input image or video sequence x, a primary encoder 710 generates a compressed representation x that is compact for storage and transmission purposes when compared to the input image x.
We can compute the compressed representation
is converted to a discrete value quantized representation using a quantizer 760.
This discrete value quantization representation
may be entropy coded into a bitstream using an arithmetic coder 770 using arithmetic coding (lossless or lossy). At the decoder side, the discrete-value quantized representation
Using an arithmetic decoder 780, the bitstream can undergo entropy decoding with lossless or lossy compression to recover this discrete-value quantized representation
to the main decoder 720 to generate the input image or video sequence
can be restored and/or reconstructed. The primary encoder 710 and the primary decoder 720 may be neural network-based encoders and decoders (e.g., DNN-based coders).

According to some embodiments, the E2E NIC framework can include a hyper-prior model and a context model during the online training phase to further improve the compression performance. The hyper-prior model can be used to capture spatial dependencies in the latent representations generated between layers of the neural network. According to some embodiments, side information may be used by the hyper-prior model, which is typically generated by motion-compensated temporal interpolation of neighboring reference frames at the decoder side. This side information can be used to train and infer the E2E NIC framework. The hyper-encoder 730 uses a hyper-prior neural network-based encoder to generate the compressed representations.
may be encoded as x′. A quantizer 761 and an arithmetic coder 771 may then be used to generate a hyper-compressed representation of the hyper-encoded compressed representation. An arithmetic decoder 781 may decode the hyper-compressed representation. A hyper-prior neural network based hyper-decoder 740 may then be used to generate a hyper-reconstructed image x′. A neural network based context model 750 may be trained using the hyper-reconstructed image and the quantized representation from the quantizer 760. The arithmetic coder 770 and the arithmetic decoder 780 may use the context model 750 to perform encoding and decoding, respectively.

According to some embodiments, the E2E NIC framework may be self-trained. The goal of the training process is to learn DNN encoding and DNN decoding (i.e., the primary encoder 710 and the primary decoder 720). In the training process, the weight coefficients of the DNN (i.e., the primary encoder 710 and the primary decoder 720) are first initialized, for example, using a pre-trained corresponding DNN model or by setting them to random values. Then, given an input training image x, the input training image x goes through the encoding process described in FIG. 4 to generate information encoded into a bitstream, and then computes the image
To reconstruct , we go through the decoding process described in Fig. 5. In the E2E NIC framework (as shown in Fig. 7), two competing goals are achieved: better reconstruction quality and lower bit consumption. The quality loss function
is used to measure the reconstruction quality (also called distortion loss), which may be Peak Signal-to-Noise Ratio (PSNR), Multi-Scale Structural Similarity (MS-SSIM), or a weighted combination of both. To measure the bit consumption of the compressed representation, rate loss
Therefore, a trade-off hyperparameter λ is used to optimize the joint rate-distortion (RD) loss according to the following equation:

Here, E measures the distortion of the decoded block residual compared to the original block residual before encoding, which acts as a regularization loss for the residual encoding/decoding DNN and encoding/decoding DNN. β is a hyperparameter that balances the importance of the regularization loss.

In some embodiments, the encoding DNN and the decoding DNN can be updated together based on backpropagated gradients in the E2E framework.

Figure 8 is a flow chart illustrating a method 800 for content-adaptive online training of an E2E NIC based on several patterns using a neural network, according to an embodiment.

In some implementations, one or more process blocks of FIG. 8 may be performed by platform 220. In some implementations, one or more process blocks of FIG. 8 may be performed by another device or group of devices that are separate from or include platform 220, such as user device 210.

As shown in FIG. 8, at operation 810, the method 800 may include receiving an input image to the E2E NIC framework.

At operation 820, the method 800 may include dividing the input image into a number of blocks.

At operation 830, the method 800 may include selecting a subset of blocks from the plurality of blocks, the subset of blocks sharing a same pattern. The same pattern may be determined based on an RGB variance of the plurality of blocks, or a YUV variance of the plurality of blocks.

At operation 840, the method 800 may include preprocessing the neural network of the E2E NIC framework. The preprocessed NN may be applied to a selected subset of blocks to fine-tune the framework.

At operation 850, the method 800 may include calculating update parameters using the preprocessed neural network. The update parameters may include a learning rate and a number of steps, where the learning rate and the number of steps are selected based on characteristics of the input image. The characteristics of the input image may be one of the RGB variance of the input image and the RD performance of the input image.

At operation 860, the method 800 may include generating an updated E2E NIC framework based on the update parameters. In some embodiments, the method further includes encoding the plurality of blocks and the update parameters to generate a compressed representation of the plurality of blocks and a compressed representation of the update parameters, decoding the compressed representation of the update parameters to generate decoded update parameters, updating the E2E NIC framework based on the decoded update parameters, and decoding the compressed representation of the plurality of blocks based on the updated E2E NIC framework to generate a reconstructed image. The distortion loss of the reconstructed image may be determined based on the consumption of the compressed representation of the plurality of blocks and the update parameters, the trade-off hyper-parameter, and the distortion between the block residual of the compressed representation of the plurality of blocks and the block residual of the decoded compressed representation of the plurality of blocks.

Although FIG. 8 illustrates example blocks of a method, in some implementations, the method may include additional, fewer, different, or differently arranged blocks than the blocks depicted in FIG. 8. Additionally or alternatively, two or more of the blocks of the method may be performed in parallel.

FIG. 9 is a block diagram of an example of computer code 900 for content-adaptive online training of an E2E NIC based on several patterns using a neural network, according to an embodiment. In an embodiment, the computer code may be, for example, program code or computer program code. According to an embodiment of the present disclosure, an apparatus/device may be provided that includes at least one processor having a memory that stores the computer program code. The computer program code, when executed by the at least one processor, may be configured to perform any number of aspects of the present disclosure.

As shown in FIG. 9, computer code 900 includes receiving code 910, dividing code 920, selecting code 930, preprocessing code 940, computing code 950, and generating code 960.

The receiving code 910 is configured to cause at least one processor to receive an input image into the E2E NIC framework.

The partitioning code 920 is configured to cause at least one processor to partition the input image into multiple blocks.

The selection code 930 is configured to cause the at least one processor to select a subset of blocks from the plurality of blocks, the subset of blocks sharing a same pattern. The same pattern may be determined based on an RGB variance of the plurality of blocks or a YUV variance of the plurality of blocks.

The preprocessing code 940 is configured to cause at least one processor to preprocess a neural network of the E2E NIC framework, and the preprocessed NN is applied to a selected subset of blocks to fine-tune the framework.

The computing code 950 is configured to cause the at least one processor to calculate update parameters using the preprocessed neural network. The update parameters may include a learning rate and a number of steps, the learning rate and the number of steps being selected based on characteristics of the input image. The characteristics of the input image may be one of the RGB variance of the input image and the RD performance of the input image.

The generating code 960 is configured to cause the at least one processor to generate an updated E2E NIC framework based on the update parameters. The computer code 900 may further include code configured to cause the at least one processor to encode the plurality of blocks and the update parameters to generate a compressed representation of the plurality of blocks and a compressed representation of the update parameters, code configured to cause the at least one processor to decode the compressed representation of the update parameters to generate decoded update parameters, code configured to cause the at least one processor to update the E2E NIC framework based on the decoded update parameters, and code configured to cause the at least one processor to decode the compressed representation of the plurality of blocks based on the updated E2E NIC framework to generate a reconstructed image.

Although FIG. 9 shows example blocks of code, in some implementations an apparatus/device may include additional, fewer, different, or differently arranged blocks relative to those shown in FIG. 9. Additionally or alternatively, two or more of the blocks of an apparatus may be executed in parallel. In other words, although FIG. 9 shows separate blocks of code, the various code instructions need not be separate and may be intermixed.

The content-adaptive online training method and process of the E2E NIC framework based on several patterns described in this disclosure provides flexibility to the adaptive online training mechanism to improve NIC coding efficiency and support different types of learning-based quantization methods, including DNN-based or traditional model-based methods. The described method further provides a flexible and generic framework that accommodates different DNN architectures and multiple quality criteria.

The techniques described above may be implemented using computer-readable instructions, as computer software physically stored on one or more computer-readable media, or by one or more tangibly configured hardware processors. For example, FIG. 2 illustrates an environment 200 suitable for implementing various embodiments. In one example, one or more processors execute a program stored on a non-transitory computer-readable medium.

As used herein, the term component is intended to be broadly interpreted as hardware, software, or a combination of hardware and software.

It will be apparent that the systems and/or methods described herein may be implemented in different forms of hardware, software, or a combination of hardware and software. The actual dedicated control hardware or software code used to implement these systems and/or methods is not intended to limit the implementation. Thus, the operation and behavior of the systems and/or methods are described herein without reference to specific software code, and it will be understood that software and hardware may be designed to implement the systems and/or methods based on the description herein.

Computer software can be encoded using any suitable machine code or computer language that can be subjected to mechanisms such as assembly, compilation, linking, etc. to create code containing instructions that can be executed by a computer central processing unit (CPU), graphics processing unit (GPU), etc., directly or through interpretation, microcode execution, etc.

The instructions may be executed on various types of computers or computer components, including, for example, personal computers, tablet computers, servers, smartphones, gaming consoles, Internet of Things devices, etc.

While this disclosure describes some exemplary embodiments, there are modifications, substitutions, and various alternative equivalents that fall within the scope of this disclosure. Thus, it will be appreciated that those skilled in the art can devise numerous systems and methods not explicitly shown or described herein that embody the principles of this disclosure and thus are within the spirit and scope of this disclosure.

200 Environment 210 User Device 220 Platform 222 Cloud Computing Environment 224 Computing Resources 224-1 Application (APP)
224-2 Virtual Machine (VM)
224-3 Virtualized Storage (VS)
224-4 Hypervisor (HYP)
230 network 300 device 310 bus 320 processor 330 memory 340 storage component 350 input component 360 output component 370 communication interface 400 image 410 block 510 subset 520 subset 700 E2E NIC framework 710 primary encoder 720 primary decoder 730 hyperencoder 740 hyperdecoder 750 context model 760 quantizer 761 quantizer 770 arithmetic coder 771 arithmetic coder 780 arithmetic decoder 781 arithmetic decoder 800 method for content adaptive online training of E2E NIC 810-860 operation 900 computer code 910 receiving code 920 splitting code 930 selection code 940 preprocessing code 950 Computing Code 960 Generated Code

Claims (7)

1. A method of content-adaptive online training for end-to-end (E2E) neural image compression (NIC) using a neural network, executed by at least one processor, the method comprising:
receiving an input image to an E2E NIC framework;
Dividing the input image into a number of blocks;
selecting a subset of blocks from the plurality of blocks, the subset of blocks sharing a same pattern;
preprocessing a neural network of the E2E NIC framework and applying the preprocessed neural network to the selected subset of blocks ;
calculating update parameters using the preprocessed neural network;
generating an updated E2E NIC framework based on the updated parameters ;
encoding the plurality of blocks and the update parameters to generate a compressed representation of the plurality of blocks and a compressed representation of the update parameters;
decoding the compressed representation of the update parameters to generate decoded update parameters;
updating the E2E NIC framework based on the decoded update parameters;
decoding the compressed representation of the plurality of blocks based on the updated E2E NIC framework to generate a reconstructed image;
determining a distortion loss of the reconstructed image based on consumption of the compressed representations of the blocks and the update parameters, a trade-off hyper-parameter, and a distortion between block residuals of the compressed representations of the blocks and block residuals of the decoded compressed representations of the blocks;
A method comprising: The method of claim 1, wherein the same pattern is determined based on an RGB variance of the plurality of blocks or a YUV variance of the plurality of blocks. The method of claim 1, wherein the update parameters include a learning rate and a number of steps, the learning rate and the number of steps being selected based on characteristics of the input image. The method of claim 3 , wherein the characteristic of the input image is one of an RGB variance of the input image and an RD performance of the input image. The method of claim 1, wherein when preprocessing the neural network, the neural network is fine-tuned using the plurality of blocks. An apparatus for content-adaptive online training for end-to-end (E2E) neural image compression (NIC) using neural networks, the apparatus executing the method according to any one of claims 1 to 5 . 6. A computer program comprising instructions, when executed by at least one processor of an apparatus for end-to-end (E2E) neural image compression (NIC) content-adaptive online training using neural networks, that cause the at least one processor to perform the method of any one of claims 1 to 5 .